Hybrid-cloud agentic applications allocate retrieval, inference, forecasting, and operational analytics across private clusters and public providers whose delay, price, carbon exposure, capacity, and policy eligibility vary over time. Optimizing mean cost or mean latency alone can hide rare but consequential deadline misses, unsupported recommendation routes, privacy-boundary violations, and demand forecast failures. This paper proposes TCRA, a Tail-Risk Controlled Resource Allocation architecture in which demand, placement, policy, carbon, and verification agents jointly construct feasible workload routes under conditional value-at-risk (CVaR) budgets. TCRA combines contract-typed actions, API-mesh telemetry, distributional loss estimates, long-horizon demand forecasts, and an immutable decision ledger. A simulated evaluation over four hybrid-cloud workload families reports that TCRA reduces the CVaR of composite service loss by 39.8% relative to policy-constrained mean optimization, reduces P99 deadline violations from 4.8% to 1.7%, and preserves zero simulated policy-invalid placements while using 8.6% less carbon exposure than latency-first allocation.
Bitla et al. (Tue,) studied this question.